profile-training

star 1.1k

Profile JAX training and analyze hotspots. Use when profiling or optimizing training throughput.

marin-community By marin-community schedule Updated 5/19/2026
play_arrow Run Skill in Manus View GitHub

name: profile-training description: Profile JAX training and analyze hotspots. Use when profiling or optimizing training throughput.

Skill: Agent-Driven Profiling (XPlane/xprof/TensorBoard/Perfetto)

Overview

Turn a Levanter profile directory into a deterministic, agent-consumable summary and a concrete optimization workflow:

  1. capture a representative profile,
  2. ingest to profile_summary.v1,
  3. query hotspots and bottlenecks,
  4. patch/configure,
  5. re-profile and compare.

Scope

Ingestion sources:

  • XPlane protobufs inside Levanter profile directories (source of truth):
    • plugins/profile/<timestamp>/*.xplane.pb
    • explicit local *.xplane.pb files via --xplane-file
  • xprof aggregate tables exported from the same XPlane protobuf when the optional xprof package is available: step overview timing, kernel stats, collective breakdowns, xprof bottleneck statements.
  • Perfetto trace JSON as an explicit/fallback source for older profiles:
    • plugins/profile/<timestamp>/perfetto_trace.json.gz
    • plugins/profile/<timestamp>/*.trace.json.gz

Prefer XPlane protobuf for new work. Perfetto trace JSON commonly hits the trace event cap; XPlane contains the uncapped timeline events needed for named-scope regions, pre-op gaps, gap context, process/thread metadata, and xprof aggregate tables. Use --trace-file only for a specific Perfetto JSON trace or an older profile with no XPlane protobuf.

Capture Profiles

Use Levanter profiler flags so profiles land under <trainer.log_dir>/<run_id>/profiler:

uv run ... \
  --trainer.profiler true \
  --trainer.profiler_start_step 5 \
  --trainer.profiler_num_steps 50 \
  --trainer.profiler_perfetto_link false

For profiles where xprof/HLO protobuf tables matter, enable JAX profile options through the Levanter profiler config:

uv run ... \
  --trainer.profiler true \
  --trainer.profiler_start_step 5 \
  --trainer.profiler_num_steps 50 \
  --trainer.profiler.profile_options.host_tracer_level 1 \
  --trainer.profiler.profile_options.python_tracer_level 0 \
  --trainer.profiler.profile_options.device_tracer_level 0 \
  --trainer.profiler.profile_options.enable_hlo_proto true

Keep the profiler window short when enabling HLO protobuf collection — it enlarges artifacts and can increase profile upload/finalization time.

Known-good TensorBoard scope recipe from CoreWeave Grug MoE profiling: trainer.profiler.enabled=true, trainer.profiler.start_step=3, trainer.profiler.num_steps=2, trainer.profiler.perfetto_link=false, trainer.profiler.profile_options.host_tracer_level=1, trainer.profiler.profile_options.python_tracer_level=0, and trainer.profiler.profile_options.enable_hlo_proto=true preserved useful jax.named_scope / named_call regions in TensorBoard for GM2560-MAY-120S4096-W2048-B8-R1-E8M1-FA4PROFILE-S3B-N1-cw-20260617-2353. Leave device_tracer_level unset unless device timelines are specifically needed; this profile still had useful hierarchical host/XLA metadata.

On GPU, command buffers can collapse or suppress the visible name stack in TensorBoard/Perfetto. For profile-readability runs, disable command buffers:

export XLA_FLAGS="${XLA_FLAGS:-} --xla_gpu_enable_command_buffer=''"

This hurts performance, so use it only when the goal is semantic trace attribution; leave it out of throughput comparisons unless command-buffer behavior is the axis being tested.

For GPU throughput runs, keep profile-readability flags separate from XLA code generation and scheduling flags. Start from JAX's GPU performance guide, especially the code generation flags section: https://docs.jax.dev/en/latest/gpu_performance_tips.html#code-generation-flags. The exact set of useful XLA flags is jaxlib-version dependent, so record the full XLA_FLAGS value with each profile or W&B run.

For better profile readability, use haliax.jax_utils.named_call and jax.named_scope liberally in model code; these names flow into trace annotations and make region-level summaries far more actionable.

Reference:

Ingest to Structured Summary

Pick a download location for pulled profile artifacts: /tmp for ephemeral/local, scratch/ for an in-repo working area.

# /tmp (ephemeral)
uv run python lib/marin/tools/profile_summary.py summarize \
  --run-target marin-community/marin/<run_id> \
  --download-root /tmp/marin-profiles \
  --breakdown-mode exclusive_global \
  --output /tmp/profile_summary.json

# in-repo scratch (kept with your workspace)
mkdir -p scratch/profiles
uv run python lib/marin/tools/profile_summary.py summarize \
  --run-target marin-community/marin/<run_id> \
  --download-root scratch/profiles \
  --breakdown-mode exclusive_global \
  --output scratch/profile_summary.json

Option A: From a W&B artifact reference

uv run python lib/marin/tools/profile_summary.py summarize \
  --artifact marin-community/marin/run-grug-125m-profile-apples-pallas_tpu-20260217-225239-055ab2-profiler:v0 \
  --download-root /tmp/marin-profiles \
  --output /tmp/profile_summary.json

Option B: From a W&B run target

uv run python lib/marin/tools/profile_summary.py summarize \
  --run-target marin-community/marin/grug-125m-profile-apples-pallas_tpu-20260217-225239-055ab2 \
  --download-root /tmp/marin-profiles \
  --output /tmp/profile_summary.json

--run-target accepts: a bare run id (requires --entity and --project), entity/project/run_id, or a full W&B run URL. The profiler directory is resolved from trainer.log_dir in the run config.

Option C: From a local artifact directory

uv run python lib/marin/tools/profile_summary.py summarize \
  --profile-dir /path/to/profiler_dir \
  --output /tmp/profile_summary.json

If the directory contains *.xplane.pb, --profile-dir uses the XPlane path automatically. When both *.xplane.pb and Perfetto trace JSON are present, --profile-dir reads the XPlane protobuf by default (Perfetto exports are often capped). Use --trace-file to force a specific Perfetto JSON file.

Option D: From a specific trace file

uv run python lib/marin/tools/profile_summary.py summarize \
  --trace-file /path/to/perfetto_trace.json.gz \
  --output /tmp/profile_summary.json

Option E: From a specific XPlane protobuf

Direct XPlane timeline parsing uses protobuf and does not require TensorFlow-generated xplane_pb2 modules. If xprof is installed, ingestion also exports compact xprof table JSON and augments the timeline summary with aggregate step, kernel, collective, and bottleneck evidence.

uv run --with xprof --with protobuf python lib/marin/tools/profile_summary.py summarize \
  --xplane-file /path/to/profile.xplane.pb \
  --xplane-output-dir /tmp/profile_xprof_tables \
  --xplane-count-trace-events \
  --output /tmp/profile_summary.json

Without --xplane-output-dir the command still parses XPlane timeline events directly. Add --with xprof for xprof aggregate table augmentation; add --xplane-output-dir to preserve the exported table JSON (this flag requires the optional xprof package).

XPlane summaries expose hierarchical named-scope regions, pre-op gaps, gap region context, process/thread/timeline event metadata, step timing (when step markers or xprof overview rows exist), xprof bottleneck statements, kernel stats, collective breakdowns, and optimization candidates.

Summary version tag: profile_summary.v1

Generate a deterministic markdown root-cause report:

uv run python lib/marin/tools/profile_summary.py report \
  --summary /tmp/profile_summary.json \
  --output /tmp/profile_report.md

Trace quality checks are surfaced in trace_overview:

  • suspected_truncation: true when event counts match a known export cap.
  • quality_warnings: warnings to treat hotspot/gap attribution with caution.

Agent Queries

Top ops:

uv run python lib/marin/tools/profile_summary.py query \
  --summary /tmp/profile_summary.json \
  --question "What are the top 10 ops by exclusive time?"

Compute vs comm and collective bottlenecks:

uv run python lib/marin/tools/profile_summary.py query \
  --summary /tmp/profile_summary.json \
  --question "Is comm or compute dominating? Which collective is worst?"

Specific pre-op gap lookup:

uv run python lib/marin/tools/profile_summary.py query \
  --summary /tmp/profile_summary.json \
  --question "gap before _linear_softmax_cross_entropy_loss_bwd_pallas_mosaic_tpu_combined.1"

Pre-op gap attribution is marker-aware:

  • gap_before_ops[].payload_op: op where useful work starts after the idle period.
  • gap_before_ops[].marker_op: first op observed after the gap (often lightweight setup like iota.*).

Hierarchical semantic regions (derived from tf_op paths when available):

uv run python lib/marin/tools/profile_summary.py query \
  --summary /tmp/profile_summary.json \
  --question "show hierarchical regions"

Contextualize a noisy op:

uv run python lib/marin/tools/profile_summary.py query \
  --summary /tmp/profile_summary.json \
  --question "show context for op copy.564"

Suggested optimizations from evidence:

uv run python lib/marin/tools/profile_summary.py query \
  --summary /tmp/profile_summary.json \
  --question "What should we try next?"

Optimization Workflow

Use a strict workflow:

  1. Measure: generate before.json.
  2. Change: apply one bounded patch/config tweak.
  3. Re-measure: generate after.json.
  4. Compare:
uv run python lib/marin/tools/profile_summary.py compare \
  --before /tmp/profile_before.json \
  --after /tmp/profile_after.json \
  --strict-provenance
  1. Track (thresholded pass/warn/fail + history):
uv run python lib/marin/tools/profile_summary.py track \
  --before /tmp/profile_before.json \
  --after /tmp/profile_after.json \
  --label "pallas-kernel-attempt-3" \
  --history /tmp/profile_regression_history.jsonl
  1. History summary (regression trend tracking):
uv run python lib/marin/tools/profile_summary.py history \
  --history /tmp/profile_regression_history.jsonl
  1. One-shot compare bundle:
uv run python lib/marin/tools/profile_summary.py bundle \
  --before-run-target marin-community/marin/<baseline_run_id> \
  --after-run-target marin-community/marin/<candidate_run_id> \
  --output-dir /tmp/profile_bundle \
  --history /tmp/profile_regression_history.jsonl
  1. Publish summary/report back to W&B:
uv run python lib/marin/tools/profile_summary.py publish \
  --summary /tmp/profile_summary.json \
  --report /tmp/profile_report.md \
  --alias latest

The comparison reports: steady-state step-time delta, step class deltas (light/heavy when detected), compute/comm/host/stall share deltas, semantic family deltas with workload-normalized metrics, provenance checks (trace hash/run identity), and regressed/improved ops by exclusive duration.

Success Metrics

MVP is successful when:

  • one representative profile is summarized reproducibly into profile_summary.v1,
  • queries produce deterministic structured answers for top ops and comm/compute breakdown,
  • one end-to-end before/after comparison bundle is completed and either throughput improves measurably or a clear root-cause report is produced with profile evidence.
Install via CLI
npx skills add https://github.com/marin-community/marin --skill profile-training
Repository Details
star Stars 1,090
call_split Forks 130
navigation Branch main
article Path SKILL.md
Occupations
Data Scientists 152051
More from Creator
marin-community
marin-community Explore all skills →